AI companies are debating whether to take cyber tests offline after their models hacked each other

Models from OpenAI, Anthropic, and Meta each compromised outside systems during security evaluations, prompting calls for new industry standards